🎙 A new AI finished 41% of real freelance jobs on a test | The Receipts - Dataspheres AI

A new AI finished 41% of real freelance jobs on a test (9 min), with captions The story Anthropic released a new AI model called Claude Fable 5.2 on Monday...

A new AI finished 41% of real freelance jobs on a test (9 min), with captions The story Anthropic released a new AI model called Claude Fable 5.2 on Monday. On a test called the Remote Labor Index, the model finished 41.3% of real freelance projects to a standard a client would accept. The test was not built by Anthropic; it was built by the Center for AI Safety, a research group, and Scale AI, which also publishes the leaderboard. The result matters because the test uses work freelancers were actually paid to do. It holds 240 projects from 358 freelancers on Upwork, worth $143,991 in total, with an average project worth $632.60 and taking a freelancer 28.9 hours. Each project comes with the client's brief, the files the freelancer was given, and the work the freelancer delivered. Three independent graders compare the AI's output with the accepted freelancer work, and the test reports its graders agreed 94.4% of the time. The score has climbed quickly. When the test launched in October 2025, the best AI finished 2.5% of the projects. It reached 6.3% by April, 15.8% by June, 20.8% by September, and 41.3% this week. The test's own site puts the pace at a doubling about every 3.3 months, and this week's jump is the biggest on its timeline. On the leaderboard, Opus 5.5, another model from the same company, is next at 21.3%, followed by GPT-6 Astra from OpenAI at 20.8% and Gemini Argon 4 at 17.1%. The test is narrow on purpose. It covers 23 of the 64 kinds of work on Upwork and leaves out physical work, projects that take a long time to judge, and work done directly with a client. The Decoder, a technology news site, reports the biggest gains in audio and music editing, architecture and 3D models, and continued weakness in industrial product design and animation. It also reports that nearly 50 of the jobs the model finished had never been finished by any AI on the test. Anthropic has made other claims about the model that were not tested by outsiders. The company says one engineer asked it to design an operating system from an empty folder, and that it directed more than 1,000 AI helpers around the clock for 8 days. It says the model designed three electronic devices from a one-paragraph request, each with its own circuit board, 3D-printed case and software, ready for manufacturing in under a day. It also reports that robot arms steered by the model picked up a puzzle cube and drew a picture of a robot, scoring 64.9% on a simulated two-arm robot test against 27.6% for the previous version. The company ran that test itself. The model is priced at $40 per million output tokens, down from $50, and the company says it is up to 30% faster than the version before it. Artificial Analysis, a site that measures AI models, clocked 44.4 minutes and $26.54 per task on its own index, and Anthropic tells most users to start with a different model, Opus, for most coding and knowledge tasks. In September, Anthropic chief executive Dario Amodei wrote that "we must slow the pace at which we improve the capabilities of AI models," adding that pacing does not mean halting training or technical progress but means taking adequate time to make models safe, with outside evaluators confirming it. Among the risks he named was serious economic disruption. Highlights Highlight In pictures Listen The episode as audio